Where the AI market is talking

OpenAI just open-sourced Codex Security

I was actually discussing solutions for this with my coworkers—building white-hat security agents. It seems like openai/codex-security could simplify a lot of that, or at least provide a version of Codex that's purpose-built for security workflows. Really exciting news!

Our position on open-weights models

This take is poorly informed and poorly thought out. It's just a cynical vibe, not a cogent argument. If they wanted to squash the competitors, they'd advocate for a ban on Open-Weight models. How else are they going to prevent their release? Anthropic's position is not that, so the cynical nefarious purpose argument isn't even rhetorically true. What they say openly, that open-weight models are harder to secure is absolutely true, and anyone unwilling to acknowledge that and let it inform their own reasoning is a risk. If open-weight models keep being released near the frontier (or even worse, at the frontier), something bad will happen, and we'll have few tools to keep it from getting worse before it gets better. How bad is hard to lay a prediction on, it's perhaps lucky if it's somethi…

Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the Incident

Also notable: we have a timeline now. The agent was active within Hugging Face from Thursday 8th to Monday 13th July. The Hugging Face post about the incident - https://huggingface.co/blog/security-incident-july-2026 - followed on Thursday 16th, and OpenAI's confession - https://openai.com/index/hugging-face-model-evaluation-secur... - came on Tuesday 21st.

Which Odyssey translation wins a blind reading test?

My test ranked first Daniel Mendelsohn and second Emily Wilson. I was aware of Wilson's translation, but not of that of Mendelsohn. This has made me curious to read Mendelsohn's translation, but in general I do not like to read translations of ancient texts, because all coherent translations are far too distant from the original. To really convey the meaning of the original text you need a sequence of explanatory notes for each sentence, not a translated sentence that attempts to have a one-to-one correspondence between words. What I like to read are the bilingual editions, like those of the Loeb library, where you can read the original, but also compare it with a translation when in doubt, which is much faster than reaching for a dictionary. Moreover, the English text allows you to brows…

Anthropic publishes a practical key-recovery attack on HAWK-256

Discovering Cryptographic Weaknesses with Claude

> And a veiled pitch to real cryptanalysis researchers: "Researchers at Anthropic then spent several hundred hours learning enough cryptography research to validate the model’s claim" Many of those researchers, particularly the primary researchers and the individual(s) driving the prompts behind these big stories, have very advanced math degrees and experience. What this shows more than anything is how ML can augment expertise, the searching of solution spaces, and the connecting of dots between existing almost-there research. But also what's left out is all the time wasted pursuing dead-ends. There's an obvious publication bias at play here, though we can't know how extreme without transparency.

Canceling "Hey"

There's a coherent argument for doing so: if the law is enforced, and no one wants it to be enforced, it creates pressure to fix the law. Laws that are on the books but not enforced lead to laws meaning nothing in the extreme case. I think there's merit to both sides and the sweet spot is somewhere in the middle.

Load Bearing Memory: The Bones Remember

I want to say this out loud. The third layer of the agentic stack has been misnamed. if you call it the context layer like Harrison Chase from LangChain then what you have is a retriever, but if you call it evidentiary then you can build something completely unique. Cleanly, there is convergence here across projects that are unrelated. This is signalling that the need is real to have a witness present in the agentic layer. I have shipped both parts together with lotor and now you can create what i am calling load-bearing memories. This allows it to display lived experience through it's evidentiary learning. Think like a guild. you go and train somewhere, create load on your self and is measured in the bones. That has become part of the entity that went through the training. it is it throu…

7.1 Earthquake in Japan

> For example to replicate a gas burner you have to have above 15KW continious power. Where are you getting this from? 15K BTU sure, but 15K watts? Do you mean the entire stove?

7.1 Earthquake in Japan

Electrical is inadequate power delivery-wise and requires completely different cookware to the point that it's a different cuisine. For example to replicate a gas burner you have to have above 15KW continious power. Good luck wiring that into your apartment, or feeding this with local solar. Or even finding an induction stovetop that won't melt its insides in a year or two at a measly 5KW.

AI revenues are growing fast, but not fast enough

You really see the first mover disadvantages in US AI labs. And how second movers (open Chinese labs) can disrupt them. First movers tend to get overcapitalized and way ahead of their skis. You saw a smaller version of this in the vector database market (remember that?) where companies like Pinecone were getting billion+ valuations for capabilities which now seem like a commodity. There's vector DB companies now with much less debt / VC obligations that don't need to clear insane revenue levels to justify investment.

Fast Remediation Is the New Trust Model (JFrog and OpenAI Zero-Day Findings)

So they are the proxy in the hugging face hacking incident? Way to bury that lede.

The Drift Principle: Coherence is not contact [pdf]

Hugging Face is being used to easily undress women and children

Hugging Face Has a Deepfake Nudes Problem

We have proof automation now

Proof automation will make things a lot easier. I'm not entirely sure that it's the only thing required, but it's a big part of the thing required. What's more interesting to me is, how do we prove that the business intent is accurately encoded in the code? This is a different problem. Not "is the logic internally coherent," but "is the logic going to give us what we want?" And how do we make sure that the logic keeps giving us what we want, iteration after iteration? I've open source a tool, Drift, for this purpose. So far, biased and anecdotally as its builder, this has been the missing layer. It works because the spec is treated as a duplicate of the code, written and committed in the repo and kept in sync with hashes. The LLM then checks for diffs manually. There are ways to use perpl…

Did AI Hack HuggingFace?

LiveOverflow posted this video on YouTube attempting to reverse engineer the potential exploit chain GPT used to hack HuggingFace. The title is the original one on YouTube. Original description from the video: An OpenAI agent reportedly escaped its sandbox, found multiple zero-days, and hacked Hugging Face... all to cheat on a cybersecurity benchmark? The story sounded almost too crazy to be true. Mohan (S1r1u5) investigated and reconstructed the likely attack chain, examined the patches, and reproduced vulnerabilities that match the public disclosures. Was this really a rogue AI, clever marketing, or "just" an agent that lost track of its task and caused real-world damage?

What Even Are Microservices?

This is what people always say but it's not really the only case, and it's not even a particularly good case. You could achieve the same organizational thing through modules, for example. The real—and the one thing you can't replicate—is data isolation. Of course there are other benefits, but it's data isolation is the differentiator of micro-services and other types of architecture.

Hugging Face CEO shares his demands of OpenAI after 'rogue' agent hack

Our position on open-weights models

I case you're wondering, HF = Hugging Face